Study Of Speech Recognition Algorithm Based On HMM And Artificial Neural Network

This thesis analyzed the existing speech recognition technology, and studied the basic theory of speech recognition, including the mathematical models generated by voice signal, preprocessing, the endpoint detection and extracting of features.On this basis, discussed the major technique which the speech recognition system realizes, with emphasis on the HMM model and self-organizing neural network (SOFM) principle and its application in speech recognition has been studied, analyzed and compared the characteristics of their identification and application of characteristics, proposed mixed model principle and the algorithms which are based on CDHMM and the SOFM, This model apply time neat network to produce and so on Uygur's phonetic feature vectors, and added to the SOFM classifier for speech recognition, The HMM-ANN model not only has HMM to the dynamic time series greatly strengthened modelling ability,but also has ANN greatly strengthened static classification ability.This article has carried on the isolated digit and the continual digit speech recognition system's experiment simulation under the C++ environment to the HMM model and the HMM-ANN model's algorithm. The result indicated that compares with the HMM, model method, the HMM-ANN model raised the speech recognition system's rate of accuracy, manifests the improvement model fully the feasibility and the validity, finally had pointed out this article studies the direction which in the future will improve.